Quota, Unquote

Do seats reserved for women change how local governments govern? In the linked Uttar Pradesh and Rajasthan samples, women-reserved seats are associated with essentially no difference in the Panchayat Advancement Index (PAI) Good Governance score, a 0 to 100 index of whether the panchayat held its meetings, filed its plan and accounts, and published its beneficiary and works lists. The UP estimate is 0.07 points, with a 95% confidence interval from -0.22 to 0.36. The Rajasthan estimate is -0.05 points, with an interval from -0.59 to 0.49. In Mumbai, where the city council's women's seats are drawn by lot, residents rate councillors in reserved wards 0.07 control-group standard deviations higher on a 14-item index, with an interval from -0.10 to 0.24 and a randomization p-value of 0.44.

The UP and Rajasthan estimates are conditional associations: the official seat-allocation mechanisms have not yet been shown to be random within the reconstructed comparison groups. The Mumbai contrast is causal, because the draw is a lottery, but its outcome is what residents perceive, not service delivery measured independently. None of these outcomes measures leakage, bribery, or corruption.

Research design

The treatment is whether the seat (sarpanch, pradhan, or Mumbai ward councillor) was reserved for a woman, not the winner's sex. The analysis compares units within the smallest reconstructed reservation-allocation strata, or, for Mumbai, within the lottery pool:

State	Reservation cycle	Comparison strata	Status
Uttar Pradesh	2021	District, LGD block, and caste-reservation class	Frozen before the outcome regression
Rajasthan	2020	Panchayat Samiti and caste-reservation class	Exploratory
Mumbai (BMC)	2007, 2012, 2017	The lottery pool: the council, and for 2012 the wards not reserved in 2007	Specified after one exploratory pass elsewhere

The primary estimator is an unweighted regression with assignment-stratum fixed effects and HC2 standard errors. Block-clustered CR2 intervals and fixed-count randomization tests are robustness checks. The design record preserves the UP specification, explains why the Rajasthan analysis is exploratory, and states the Mumbai specification and what it was blind to. Commit edd8b42 records the UP freeze.

Data

The project joins four sources:

• PAI Good Governance scores for fiscal years 2022-23 and 2023-24.
• Uttar Pradesh local elections, including the 2021 reservation category and reviewed LGD links.
• Rajasthan reservation data for the 2020 sarpanch cycle.
• local_reservations for Mumbai: the seat reservation of the 2007, 2012 and 2017 councils and the Praja Foundation's ward-level citizen ratings of councillors, six survey waves from 2011 to 2018, mirrored there from the CC0 replication deposit of Karekurve-Ramachandra and Lee (2025), doi:10.7910/DVN/IO9SLQ.

PAI 2.0 Good Governance for 2023-24 is the primary outcome. The Ministry of Panchayati Raj scores each Gram Panchayat on nine themes; Good Governance (theme 8) is built from equally weighted indicators that the panchayat reports on the PAI portal and the Gram Sabha and district validate. In PAI 2.0, 23 of its 26 indicators are yes/no checks: was a Gram Sabha, a Mahila Sabha and a Bal Sabha held; do standing committees meet; was the development plan uploaded to eGramSwaraj by 31 March; were provisional accounts closed within 15 days of year end and read out in the Gram Sabha; are beneficiary and works lists displayed and approved; is there a grievance system, online payment, online services, a co-located common service centre, GeM procurement, a disaster plan; is the office open. The other three are ratios: growth in own-source revenue, and the shares of planned activities initiated and completed. So the primary outcome mostly records whether required procedures were carried out and documented, and a null means women-reserved panchayats are no more or less likely to hold the meetings, file the plan, close the books, publish the lists, and grow their own revenue.

Version	Fiscal year	Indicators	Ratios	Yes/no checks	Shared with the other version
PAI 1.0	2022-23	62	27	35	10
PAI 2.0	2023-24	26	3	23	10

PAI 1.0's theme 8 was closer to performance: a share of grievances redressed, of services delivered within the Citizens' Charter time, of works completed and geo-tagged, of issues raised by SC/ST, women, elderly and disabled residents acted on, and whether the social audit report was uploaded. A ratio has a denominator distinct from its numerator on the portal, whatever the label says; a check is a yes/no question. PAI 2.0 dropped most of these ratios for compliance checks, and only 10 indicators appear in both, so the 2022-23 score is a separate replication rather than a second observation of the same outcome. The full lists, fetched from the portal's indicator browser, are in docs/pai_theme8_indicators.csv.

Both PAI vintages carry LGD Gram Panchayat codes in PAI release v0.2.0, so each wave joins directly on the reviewed election-to-LGD link. Rajasthan's election panel has an LGD code for 4,729 of its 7,882 GPs; the rest use exact normalized names within manually reviewed crosswalks for reorganized districts and blocks. Fuzzy name proposals never enter the primary link; a reviewed fuzzy link could only feed a robustness variant. Failed links remain missing and are never coded as zero.

State	Election GPs	PAI 2.0 linked	Link rate	Estimation sample
Uttar Pradesh	49,773	38,388	77.1%	38,277
Rajasthan	7,882	5,723	72.6%	5,422
Mumbai (BMC)	681 ward seats	681 rated	100%	681

UP link rates are 76.8% for women-reserved seats and 77.3% for other seats. Rajasthan rates are 73.0% and 72.3%, respectively. Similar rates reduce concern about differential linkage, but they do not recover unlinked Gram Panchayats. In UP the ceiling is the election release, not PAI: 38,397 of the 49,773 winners carry a reviewed LGD code, and PAI 2.0 scores all 57,678 UP Gram Panchayats, as the official release reports.

Key results

State	Outcome	Estimate	95% CI	Control SDs	Sample
Uttar Pradesh	PAI 2.0	0.07	[-0.22, 0.36]	0.004	38,277
Rajasthan	PAI 2.0	-0.05	[-0.59, 0.49]	-0.004	5,422
Mumbai (BMC)	Praja 14-item rating index	0.05	[-0.08, 0.18]	0.067	681

The UP confidence interval excludes improvements larger than 0.020 control-group standard deviations in the linked, informative-strata sample. Its PAI 1.0 estimate is -0.05 points (95% CI: -0.34, 0.24). Restricting PAI 2.0 to exact election-to-LGD links, clustering by block, and using fixed-count randomization inference all yield the same near-zero pattern. The frozen specification was first estimated on a 48% linked sample drawn from an incomplete PAI extract, giving -0.04 points (95% CI: -0.40, 0.31); the design record keeps that result beside the re-estimation.

The exploratory Rajasthan PAI 1.0 estimate is 0.25 points (95% CI: -0.26, 0.76). The underlying values are generated by the analysis scripts and stored in tabs/up_pai_effects.csv and tabs/raj_pai_effects.csv.

Mumbai's index is the mean of fourteen within-wave standardised ratings (roads, water, schools, sanitation, corruption, and so on), one survey wave per council. The interval in control-group standard deviations runs from -0.10 to 0.24, so it neither shows an improvement nor rules out the 0.16 to 0.18 that Desai, Karekurve-Ramachandra and Montero (2024) report from the same surveys plus a non-public 2019 wave and Praja's own aggregate grade. The 18-item index and dropping the one ward whose reservation the two sources dispute give the same picture. Item by item, exploratory and unadjusted, residents in reserved wards rate schools, law and order, power and water somewhat higher and the councillor's personal accessibility no differently. Councillors in reserved seats attend ward-committee meetings more (0.18 SD, 95% CI 0.01 to 0.34) and ask fewer questions in council (-0.21 SD, -0.38 to -0.04). Values are in tabs/bmc_praja_effects.csv, tabs/bmc_praja_items.csv and tabs/bmc_praja_activity.csv.

Limits of the evidence

The UP and Rajasthan comparisons are not yet causal. Reservation laws establish rotation and minimum shares, but they do not by themselves prove random allocation within the reconstructed strata. Mumbai's draw is a lottery, so its contrast is causal for the seat.

PAI cannot answer whether women leaders reduce corruption. Its Good Governance theme does ask about procurement through GeM, closing the accounts, and Gram Sabha approval of beneficiary lists, but as checks that a procedure exists, not as measures of leakage, audit findings, procurement anomalies, or beneficiary fraud. Those outcomes are required for a corruption claim.

Linkage restricts the result to the three quarters of UP Gram Panchayats with a reviewed LGD code, and to linked Rajasthan GPs. Similar coverage by treatment is useful evidence against one selection mechanism, not evidence for the missing population.

Mumbai's outcome is a perception survey of about 100 residents per ward, and one of its items, satisfaction with the councillor, is inverted in the 2018 wave and excluded. The public data lack the 2019 wave. Three councils give 681 seats, so the interval is wide: the design can tell a 0.3 SD effect from zero, not a 0.1 SD one.

Reproduce the analysis

Requirements:

• R 4.6 or newer
• Python 3.12 or newer and uv
• XeLaTeX and latexmk

git clone https://github.com/in-rolls/quota_unquote.git
cd quota_unquote
R -e "renv::restore()"
uv sync --all-groups
PAI_RELEASE_FILE=/path/to/pai_gp.parquet Rscript scripts/99_run_all.R
make paper

PAI_RELEASE_FILE must be the pinned pai_gp.parquet from PAI release v0.2.0, also on Hugging Face. The pipeline reads the Rajasthan, UP and Mumbai files from sibling repositories by default. QUOTA_RAJ_PANEL, UP_ELECTION_FILE and LOCAL_RESERVATIONS_DIR can point to other copies of the pinned files. Expected paths, source commits, and SHA-256 hashes are recorded in data/manifest.yaml.

Run the project checks separately:

make sync
PAI_DATA_DIR=/path/to/pai/consolidated make data
make check

make data rebuilds the joined data, linkage audits, estimates, tables, and figures. make check runs the R and Python tests, lint checks, design gates, and manuscript build.

Repository guide

• scripts/01* prepare and profile source data.
• scripts/02* perform audited, row-preserving joins.
• scripts/03* estimate the state results and build the comparison figure; 03d is Mumbai.
• scripts/98* enforce design and disclosure contracts.
• tabs/ and figs/ contain generated results.
• docs/design.md records analysis status and specifications.
• docs/data.md defines variables, recodes, and linkage contracts.

The manuscript source is ms/main.tex; make paper builds ms/main.pdf.